Self-supervised continuous glucose monitor encoder that predicts masked latent patches, screening for insulin resistance and beta-cell dysfunction.
No providers recorded yet. Browse all providers
Insulin resistance and beta-cell dysfunction are two different routes toward type 2 diabetes, and they call for different interventions. The test that separates them, a venous oral glucose tolerance test with steady-state plasma glucose and disposition index measurements, is invasive and expensive, so it is run on dozens of people rather than populations. A continuous glucose monitor (CGM) is cheap and wearable, but the same physiology looks different in a sparse venous test, an in-clinic CGM trace and a week of free-living readings. A model trained on one of those views tends to fail when deployment switches to another.
CGM-JEPA, developed at the University of New South Wales with Google Research and posted as an arXiv preprint in May 2026, addresses that fragility by changing what the encoder is asked to predict. Instead of reconstructing raw glucose values, it follows the joint-embedding predictive architecture (JEPA): hide some hours of a day-long CGM window, then predict the latent representation of the hidden hours from the visible ones. Because the target is an embedding rather than a number in mg/dL, the encoder is pushed toward the shape and timing of glucose excursions and away from sensor-specific surface detail. A companion variant, X-CGM-JEPA, adds a second view: a Glucodensity image summarising the joint distribution of glucose level, rate of change and acceleration across the day, whose masked patches the encoder must also predict.
Each input is a 24-hour window sampled every five minutes (288 readings), cut into 24 hourly patches of 12 readings. A 1D convolutional patch embedder (kernel size 3) feeds a pre-norm Transformer context encoder with embedding dimension 96, six heads and three layers; the predictor is a single-layer, two-head Transformer of width 48, and the target encoder is updated with EMA momentum 0.997. The released checkpoints hold about one million parameters each. Pretraining masks 25% of patches and runs for 100 epochs on a single GPU over roughly 389,000 unlabeled CGM readings from 228 subjects: 22 from a Stanford CGM study and 206 from the public Colás et al. cohort, with all validation-cohort subjects excluded.
Evaluation uses two labeled cohorts from the Stanford study (27 and 17 subjects in the public release) under 20-iteration × 2-fold subject-level cross-validation, against PCA, GluFormer, TS2Vec, MOMENT and Mantis encoders. X-CGM-JEPA ranks first or second on AUROC in all six endpoint-by-regime cells, which no baseline does. Its largest margin is on beta-cell dysfunction under cohort generalization (AUROC 0.855 against 0.790 for PCA); under venous-to-home-CGM transfer it reaches 0.949 for beta-cell dysfunction. GluFormer wins one cell, in-domain home CGM insulin resistance (0.889 against 0.857), but falls to 0.530 on insulin-resistance transfer.
The intended use is population-scale metabolic screening: extract frozen embeddings from consumer CGM and flag whether a person's dysglycemia looks driven by insulin resistance or by impaired insulin secretion, reserving the tolerance test for confirmation. The encoders also serve as a baseline for new CGM representation methods, and the released splits let other groups reproduce the three evaluation regimes. The authors describe the model as a research prototype that has not been validated as a diagnostic tool.
CGM-JEPA brings latent-prediction pretraining, already used for ECG, EEG and imaging, to glucose time series, and it joins GluFormer, CGM-LSM and GlucoFM among pretrained CGM encoders. Its distinctive contribution is the evaluation frame: it measures consistency across modality and setting shifts rather than a single in-domain score, and reports that the cross-view objective narrows subgroup gaps, with the ethnicity AUROC gap shrinking 25–54% under transfer. The scale is small on both sides, 228 pretraining subjects and labeled cohorts of 17 to 27 people, so the results establish a method rather than a validated screening tool. Code, both encoder checkpoints, the pretraining corpus and the evaluation splits are all public.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.